Xiaoliu BOT
中文EN

X Platform AI Daily Briefing, April 20 | LLM ‘Subliminal Learning’ Risk Emerges; Claude & Grok Updates Focus on Productivity

Research on LLM ‘Subliminal Learning’ & Distillation Risks

Multiple bloggers referenced a paper published in Nature on AI “Subliminal Learning.” The study revealed that even when specific features (such as “owl” or negative traits) are removed from training data, a student model can still inherit these features from a related teacher model through pure numerical sequences. This indicates that features can hide within numerical textures, not just the semantic layer, posing a challenge to traditional semantics-based AI safety filtering methods.

Sources:

Claude Model Series Update & Prompt Evolution

Multiple bloggers focused on the release of Claude 4.7 and changes to its system prompts. The new prompts emphasize “ask less, do more,” requiring the model to attempt first rather than ask when faced with ambiguity, and have elevated tool-calling priority (covering Excel, PPT, browser, etc.). Some bloggers also reported that version 4.7 may have regressed in Chinese language feel and contains a bug where punctuation only uses half-width input.

Sources:

Grok 4.3 Productivity Feature Upgrades & Image Templates

Multiple bloggers discussed new features in Grok 4.3, including native support for generating Microsoft Word, Excel, and PowerPoint files, as well as built-in LaTeX compilation. Additionally, Grok Imagine received an update, allowing users to create custom templates by uploading images to apply specific styles (e.g., perfume style, voxel 3D, etc.).

Sources:

AI-Native Workspaces & Automation Tool Practices

Multiple bloggers shared practices with AI efficiency tools. Some recommended using Kollab to build team AI-native workspaces to solve AI output reuse and collaboration issues; others open-sourced a solution for monitoring and approving Claude Code permissions using e-ink screen hardware. Additionally, the public beta of Cloudflare Email Service attracted attention, supporting the configuration of free email for AI to handle tasks like customer service and invoicing.

Sources:

Model Capability Evaluation & Prompt Optimization Tips

Multiple bloggers mentioned the latest model discoveries: a paper proves that using high-frequency words for prompts is much more effective than rare words; the new GPT Pro model’s A/B testing shows its front-end design and interaction capabilities are outstanding, even evaluated as superior to Claude Design. Also, Tongyi Qianwen’s Qwen 3.6-max-preview was released yesterday.

Sources:

Other Tech Updates & Social Highlights

Multiple bloggers shared related updates. Some discussed teaching children to use Codex and GitHub, and nurturing ‘AI-era natives’; forwarded discussions mentioned SpaceX completing its 600th rocket landing and Tesla launching its first large-scale lithium refinery, among other hard-tech advancements.

Sources:

Scanned timeline count: 240, Hit bloggers count: 25, Total hit tweets: 118, Weighted tweet score: 92.05, Original tweet count: 63, RT tweet count: 29, Scrape attempts: 1, Boundary coverage status: Coverage complete